You keep rewriting the prompt. Adding adjectives. Adding rules. Adding act as a world-class expert who… Still garbage.
The model isn’t the problem. The model has amnesia, and you keep handing it a sticky note and expecting a strategy.
Here’s what the operators actually running AI-powered businesses do differently: they stopped writing clever prompts months ago. They write context. The prompt becomes one boring line at the end of a much bigger system.
Prompt engineering vs. context engineering
Prompt engineering assumes the model is smart and just needs the right incantation. Context engineering assumes the model is a very fast intern with no memory of your business, your customers, or what you decided last Tuesday — and your job is to give it a briefing packet before it opens its mouth.
Anthropic’s own guidance on building with Claude keeps pointing at the same thing: examples, structured references, and explicit constraints beat clever wording every time. The RAG pattern exists because retrieval consistently outperforms bigger prompts.
So stop optimizing sentences. Start engineering the room the model walks into.
The three documents that do 90% of the work
Every agent or workflow I run in production is backed by three files. Not fifty. Three. If you’re missing one, output quality collapses in predictable ways.
1. The workflow doc — what and why
This is a plain-language explanation of the job. Not the prompt. The job.
- What is this agent doing in the business?
- Who is the end reader or user?
- What is a successful output? What is an unacceptable one?
- What are the inputs it will receive, in what format?
- What happens downstream — who or what consumes the output?
Example: for a LinkedIn drafting agent, my workflow doc says the agent is writing first drafts for a founder to edit, not final posts. That single sentence changes everything about how it writes — it stops trying to be polished and starts leaving room for the human.
2. The voice / decision doc — how it should think
This is where most people go wrong. They put voice inside the prompt (“write in a punchy, direct tone…”) and wonder why every post sounds like a LinkedIn parody account.
Voice is a document, not an adjective. Mine includes:
- 10-20 real examples of accepted output, with notes on why each works
- 5-10 examples of rejected output, with notes on why they failed
- A list of banned phrases (“in today’s landscape,” “unlock,” “game-changer”)
- Explicit decision rules: when to refuse, when to ask, when to just ship
I wrote more on this in Voice Prompt Is the Moat — the short version is that your voice doc is the thing competitors can’t copy, because it’s built from your actual receipts.
3. The retrieval layer — the actual receipts
This is the part solo founders skip and then blame the model. The retrieval layer is the searchable memory the agent pulls from at runtime: your past posts, your product docs, your customer notes, your pricing, your positioning, your case notes.
You don’t need a vector database on day one. Start dumber:
- A single Google Doc or Notion page the agent gets fed as context
- A structured JSON file with your positioning, offers, and ICP
- A folder of markdown files the workflow reads before running
When you outgrow that, move to a real retrieval setup — Claude’s tool use, an n8n node that queries a database, or a lightweight vector store like Supabase or Pinecone. But the pattern is the same: the agent doesn’t guess, it looks things up.
What the prompt actually looks like now
Once you have those three documents wired in, the prompt gets boring. Here’s a real one I use for a content agent:
“Using the workflow doc, voice doc, and the three most relevant past posts from retrieval, draft one LinkedIn post about the topic below. Follow the voice doc’s structure rules. Output only the post.”
That’s it. No act as. No you are a world-class. No emojis begging the model to try harder. The context does the work; the prompt just points.
How to build this in a week
You don’t need to rebuild your stack. You need to spend [X] hours on documents you should already have.
- Day 1-2: Write the workflow doc for your single most-used AI task. One page.
- Day 3-4: Build the voice/decision doc. Paste 15 real examples. Annotate them.
- Day 5: Dump your positioning, offers, and 20 best past outputs into a single retrieval file.
- Day 6: Rewrite your prompt as one boring sentence that references the three docs.
- Day 7: Run 10 outputs. Compare to your old prompt-only setup. Update the voice doc based on what still misses.
If you’re doing this inside n8n or a Claude project, load the three docs as system context or as retrieved files at the top of the run. If you’re doing it in the Claude web app, use Projects and paste them into the knowledge section.
Why this is the actual moat
Anyone can copy your prompt. Nobody can copy your voice doc, your decision rules, or your 200 annotated past outputs. That’s proprietary. That compounds. That’s the difference between a founder playing with AI and a founder running a business on it.
Before you automate anything new, get this layer right — I wrote about the order to build things in Automate In This Order, and context engineering sits underneath every workflow on that list. Skip it and you’ll end up debugging the same failures I did in the 4 n8n + Claude workflows that died in production.
Prompt engineers optimize sentences. Context engineers ship businesses. Pick one.